准备工作:
1、代码开源框架使用的是 fizyr/keras-retinanet
2、Keras版本要2.2.4以上
下面进入正题。
第一部分:模型准备
(1)模型下载地址:fizyr/keras-retinanet
(2)模型安装可以使用以下命令:
pip install numpy --user
pip install . --user
安装过程中,会检查依赖库,比如opencv-python,如果没有安装,会加载并安装。这里提一句,如果在安装时某个包下载安装不成功,自己记下来版本,比如opencv-python 3.4.5.20,可以直接先去利用pip或conda安装,但是一定要记得对应的版本。
(3)模型编译可以使用以下命令:
python setup.py build_ext --inplace
编译的时候可能会出现提示,没有某个版本C++的编译器,我提示的时没有2014版,把错误提示直接百度,就会出现解决方法,我是下载了一个3M的14版的编译工具。(当然,最好就是有相应版本的完整C++)
第二部分:数据准备
(1)在keras-retinanet-master/keras_retinanet/
文件夹下面新建一个文件夹CSV
用来存放自己制作的数据集。
数据文件夹格式如下:
|———— train_annotations.csv # 必须
|———— val_annotations.csv # 必须
|———— classes.csv # 必须
|
|____ data # (可选),这样 annotations.csv可以使用图片的相对路径
└─ *.jpg
(2)根据官网的样例,自己制作的Annotations数据集格式如下:
path/to/image.jpg,x1,y1,x2,y2,class_name
如果一张图片中没有包含任何要检测的物体,则格式如下:
path/to/image.jpg,,,,,
一个完整的例子:
/data/imgs/img_001.jpg,837,346,981,456,cow
/data/imgs/img_002.jpg,215,312,279,391,cat
/data/imgs/img_002.jpg,22,5,89,84,bird
/data/imgs/img_003.jpg,,,,,
下面,我就贴出自己写的一个代码:
def restrict_image_info(label_path):
with open(label_path, 'r') as load_f:
load_dict = json.load(load_f)
image_collect = load_dict['images']
image_num = len(image_collect)
anno_collect = load_dict['annotations']
anno_num = len(anno_collect)
img_path_list = []
x1_list = []
y1_list = []
x2_list = []
y2_list = []
category_list = []
img_path_val_list = []
x1_val_list = []
y1_val_list = []
x2_val_list = []
y2_val_list = []
category_val_list = []
mapper = {0: 'tieke', 1: 'heiding',
2: 'daoju', 3: 'dian', 4: 'jiandao'}
train_rate = 0.9
hight = image_num*train_rate
train_img_id = np.random.randint(0, image_num, size=int(hight))
print(len(train_img_id))
for i in range(image_num):
img = image_collect[i]
img_name = img['file_name']
img_id = img['id']
img_height = img['height']
img_width = img['width']
if i in train_img_id:
for j in range(anno_num):
if anno_collect[j]['image_id'] == img_id:
bbox = anno_collect[j]['bbox']
img_path_list.append(restrict_rele_path+img_name)
x1_list.append(int(np.rint(bbox[0])))
y1_list.append(int(np.rint(bbox[1])))
x2_list.append(
int(np.rint(bbox[0] + bbox[2])))
y2_list.append(
int(np.rint((bbox[1]+bbox[3]))))
category_list.append(anno_collect[j]['category_id']-1)
anno = pd.DataFrame()
anno['img_path'] = img_path_list
anno['x1'] = x1_list
anno['y1'] = y1_list
anno['x2'] = x2_list
anno['y2'] = y2_list
anno['class'] = category_list
anno['class'] = anno['class'].map(mapper)
else:
for j in range(anno_num):
if anno_collect[j]['image_id'] == img_id:
bbox = anno_collect[j]['bbox']
img_path_val_list.append(restrict_rele_path+img_name)
x1_val_list.append(int(np.rint(bbox[0])))
y1_val_list.append(int(np.rint(bbox[1])))
x2_val_list.append(
int(np.rint(bbox[0] + bbox[2])))
y2_val_list.append(
int(np.rint((bbox[1]+bbox[3]))))
category_val_list.append(
anno_collect[j]['category_id']-1)
anno_val = pd.DataFrame()
anno_val['img_path'] = img_path_val_list
anno_val['x1'] = x1_val_list
anno_val['y1'] = y1_val_list
anno_val['x2'] = x2_val_list
anno_val['y2'] = y2_val_list
anno_val['class'] = category_val_list
anno_val['class'] = anno_val['class'].map(mapper)
anno.to_csv('CSV/train_annotations.csv', index=None, header=None)
anno_val.to_csv('CSV/val_annotations.csv', index=None, header=None)
训练图片生成的数据格式如下:
data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,88,253,206,295,daoju
data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,296,244,414,344,jiandao
data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,231,239,299,341,jiandao
data/jinnan2_round1_train_20190305/restricted/190119_184244_00166940.jpg,99,278,194,320,dian
验证图片生成的数据格式如下:
data/jinnan2_round1_train_20190305/restricted/190119_182957_00166754.jpg,314,237,326,265,dian
data/jinnan2_round1_train_20190305/restricted/190127_100838_00177153.jpg,246,229,304,279,tieke
data/jinnan2_round1_train_20190305/restricted/190119_184522_00166980.jpg,668,409,717,432,dian
data/jinnan2_round1_train_20190305/restricted/190119_183142_00166782.jpg,565,326,708,432,jiandao
data/jinnan2_round1_train_20190305/restricted/190127_143529_00178527.jpg,8,262,45,326,heiding
(3)根据官网的样例,自己制作的classes数据集格式如下:
class_name,id
一个完整的例子:
cow,0
cat,1
bird,2
最后生成的数据格式如下:
tieke,0
heiding,1
daoju,2
dian,3
jiandao,4
注意:保存的csv文件是没有头部行的,不然后续代码会报错!
(4)检查生成的数据是否合格
要进行这一步,必须先要完成第一步中模型的下载与编译!
检查数据可以使用以下命令:
python keras_retinanet/bin/debug.py csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv
其中第一个参数csv
代表要检查的数据是自己制作的数据集,第二个参数是train_annotations.csv
对应的路径,第三个参数是classes.csv
对应的路径。
(5)图片存放位置
这个可以根据自己的需要定,但是最好放在上面新建的CSV
文件夹下面,这个使用路径比较方便。在我自己这个代码中,我是在CSV
文件夹下新建一个data
文件夹下存放自己的图片,此时注意与train_annotations.csv
文件中的图片路径要一致,比如我这时候就应该是这样:
data/jinnan2_round1_train_20190222/restricted/190119_185206_00167075.jpg,125,279,177,339,tieke
data/jinnan2_round1_train_20190222/restricted/190119_185206_00167075.jpg,153,363,238,549,daoju
(6)关于模型的图片输入尺寸
在https://github.com/fizyr/keras-retinanet/blob/master/keras_retinanet/bin/train.py
中的409、410行有设置输入的默认参数(800*1333):
parser.add_argument('--image-min-side', help='Rescale the image so the smallest side is min_side.', type=int, default=800)
parser.add_argument('--image-max-side', help='Rescale the image if the largest side is larger than max_side.', type=int, default=1333)
第三部分:模型训练
模型训练可以使用以下命令:
python keras_retinanet/bin/train.py csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv --val-annotations keras_retinanet/CSV/val_annotations.csv
其中第一个参数csv
代表要检查的数据是自己制作的数据集,第二个参数是train_annotations.csv
对应的路径,第三个参数是classes.csv
对应的路径,第四个参数--val-annotations
是val_annotations.csv
对应的路径。
多卡训练可用如下命令:
python keras_retinanet/bin/train.py --multi-gpu-force --multi-gpu 2 --batch-size 2 csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv --val-annotations keras_retinanet/CSV/val_annotations.csv
替换backbone可用如下命令(可选的包括vgg16,vgg19,resnet50,resnet101,densenet121,densenet169,densenet201):
python keras_retinanet/bin/train.py --steps 1000 --backbone vgg16 --gpu 2 csv keras_retinanet/CSV/train_annotations.csv keras_retinanet/CSV/classes.csv --val-annotations keras_retinanet/CSV/val_annotations.csv
第四部分:模型测试
#!/usr/bin/env python
# coding=UTF-8
'''
@Description:
@Author: HuangQinJian
@LastEditors: HuangQinJian
@Date: 2019-03-07 08:59:34
@LastEditTime: 2019-03-07 11:13:20
'''
import os
import time
import keras
# import miscellaneous modules
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
# set tf backend to allow memory to grow, instead of claiming everything
import tensorflow as tf
from skimage.io import imsave
import cv2
from keras_retinanet import models
from keras_retinanet.utils.colors import label_color
from keras_retinanet.utils.image import (preprocess_image, read_image_bgr,
resize_image)
from keras_retinanet.utils.visualization import draw_box, draw_caption
if not os.path.exists('result'):
os.mkdir('result')
def get_session():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
return tf.Session(config=config)
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
def predict_save(model, test_img_fold, test_img_list):
# load image
img_name_list = []
bboxes_list = []
class_list = []
score_list = []
for i in range(len(test_img_list)):
# for i in range(1):
img_name = test_img_list[i]
img_path = os.path.join(test_img_fold, img_name)
image = read_image_bgr(img_path)
# copy to draw on
draw = image.copy()
draw = cv2.cvtColor(draw, cv2.COLOR_BGR2RGB)
# preprocess image for network
image = preprocess_image(image)
image, scale = resize_image(image)
# process image
start = time.time()
# print(image.shape)
# print(scale)
boxes, scores, labels = model.predict_on_batch(
np.expand_dims(image, axis=0))
print("processing time: ", time.time() - start)
# correct for image scale
boxes /= scale
i = 0
for box, score, label in zip(boxes[0], scores[0], labels[0]):
# scores are sorted so we can break
if score < 0.5:
break
color = label_color(label)
b = box.astype(int)
img_name_list.append(img_name)
bboxes_list.append(b)
class_list.append(labels[0][i])
score_list.append(score)
i += 1
draw_box(draw, b, color=color)
caption = "{} {:.3f}".format(labels_to_names[label], score)
draw_caption(draw, b, caption)
imsave('result/'+img_name, draw)
submit = pd.DataFrame()
submit['img_name'] = img_name_list
submit['bbox'] = bboxes_list
submit['class'] = class_list
submit['score'] = score_list
# submit.to_csv('submit.csv', index=None)
submit.to_pickle('submit.pkl')
if __name__ == "__main__":
# set the modified tf session as backend in keras
keras.backend.tensorflow_backend.set_session(get_session())
# adjust this to point to your downloaded/trained model
# models can be downloaded here: https://github.com/fizyr/keras-retinanet/releases
model_path = os.path.join('snapshots', 'old.h5')
# load retinanet model
model = models.load_model(model_path, backbone_name='resnet50')
# if the model is not converted to an inference model, use the line below
# see: https://github.com/fizyr/keras-retinanet#converting-a-training-model-to-inference-model
model = models.convert_model(model)
# print(model.summary())
# load label to names mapping for visualization purposes
labels_to_names = {0: 'tieke', 1: 'heiding',
2: 'daoju', 3: 'dian', 4: 'jiandao'}
test_img_fold = 'keras_retinanet/CSV/data/jinnan2_round1_test_a_20190306/'
test_img_list = os.listdir(test_img_fold)
print(len(test_img_list))
predict_save(model, test_img_fold, test_img_list)
可能会遇到的错误:
(1)ImportError: No module named 'keras_resnet'
解决办法:pip install keras-resnet --user
(2)在第四部分模型预测的时候,必须运行:
# if the model is not converted to an inference model, use the line below
# see: https://github.com/fizyr/keras-retinanet#converting-a-training-model-to-inference-model
model = models.convert_model(model)
否则会报如下错误:
'boxes, scores, labels ' not enough values to unpack (expected 3, got 2)
参考资料:
1、Retinanet训练自己的数据(2):模型准备